Why Youth Homelessness Intervention Evidence Matters
Youth homelessness is a complex and evolving issue. Young people experiencing homelessness or housing instability may interact with many systems, including housing, education, health care, child welfare, justice, and workforce services. Interventions may focus on housing, outreach, case management, family support, mental health, education, employment, sexual health, or other areas of well-being.
Because the field is broad, researchers, service providers, policymakers, students, and community partners need evidence that is timely and usable. They need to know what interventions have been tested, what outcomes they improve, for whom they work, and where evidence remains limited.
Systematic reviews help answer these questions by bringing together findings across many studies. But they are also slow and labor-intensive. Searching databases, screening titles and abstracts, reviewing full texts, extracting data, and synthesizing findings can take months or years. By the time a review is completed, new studies may already be available.
That is why this project focuses on living evidence: evidence that can be revisited, updated, and made easier to use over time.
This prototype builds on a prior systematic review of interventions for homeless and unstably housed youth. That review focused on youth ages 13 to 25 who were homeless, unstably housed, runaway, or at risk of homelessness, and it examined interventions in OECD country contexts. The protocol included two main goals: synthesizing evidence on intervention effectiveness and understanding implementation factors related to why interventions may succeed or fail.
The original human-led review searched studies from 2008 to 2017 and was later published as both a Chapin Hall policy report and a peer-reviewed journal article:
This earlier review created a valuable baseline: a set of human-screened studies and eligibility decisions that could be used to develop and evaluate an AI-assisted screening workflow.
Building on an Existing Evidence Base
The earlier review was an important evidence product, but like most systematic reviews, it represented a completed search window. Once the review was finished, the field continued to move. New studies appeared after 2017, and updating the evidence base would require repeating parts of the review process.
This prototype reimagines the review as a living systematic review supported by AI.
The goal is not to replace human reviewers. Instead, the project asks how AI can help make one part of the review process faster, more transparent, and easier to update.
Abstract screening is an ideal starting point because it is early, repetitive, and consequential. If relevant studies are excluded too early, they may never reach full-text review. If too many irrelevant studies are retained, the review becomes slower and more difficult to manage.
In this project, AI is treated as a human-guided screening assistant. Human decisions remain the benchmark for validation, interpretation, and final judgment.
From Static Review to AI-Assisted Living Review
A Two-Phase Prototype
This project was developed in two phases.
Phase 1: Prompt Development Using Baseline Review Data
In Phase 1, the project used the human-screened studies from the prior review period, 2008–2017, as baseline data. These data were used to develop, test, and compare different AI prompts for abstract screening.
The purpose of Phase 1 was to answer:
How should the AI screening task be framed?
Should AI behave like a final reviewer or a triage assistant?
Which prompt improves recall while still producing useful screening decisions?
What kinds of abstracts are missed under different prompt designs?
How closely do AI screening decisions align with human screening decisions?
The best-performing prompt from Phase 1 was then selected for testing in the next phase.
Phase 2: Testing the Best Prompt on Newer Studies
In Phase 2, the best prompt from Phase 1 was applied to newer studies from 2018–2025. This phase tests whether the AI-assisted screening workflow can support an updated evidence review beyond the original published review period.
This second phase is especially important because it moves the project from prompt development to real-world updating. It asks whether AI can help extend a completed systematic review into a more living review process.
The AI component matters because it changes the review from a static evidence product into a more updateable workflow.
AI-assisted screening can help by:
reducing the burden of reviewing large numbers of abstracts;
making screening logic more explicit through prompt design;
allowing researchers to compare AI decisions with human decisions;
identifying where the model misses relevant studies;
supporting faster updates when new studies are published;
creating a reproducible workflow that other research teams can adapt.
The most important lesson is that AI-assisted review is not simply about automation. It is about designing a transparent human-in-the-loop process where AI supports screening, but humans remain responsible for methodological decisions, validation, and interpretation.
Why the AI Component Matters
What This Website Shows
This website documents the prototype process. It explains how the original evidence base was used to develop AI prompts, how the best prompt was selected, and how that prompt was applied to newer studies.
Visitors can use the site to explore: